An Internet-based Music Generation System using Reinforcement Learning

Abstract

Deep learning has been widely applied to digital art and music creation. However, producing melodies that follow music theory and match human compositional patterns remains challenging. This study proposes a symbolic music generation system that integrates supervised learning and reinforcement learning. The core framework employs recurrent neural networks for sequence modeling, while a reinforcement learning module formulates music rules as reward functions to guide pitch, duration, and rhythm. This hybrid approach helps produce outputs that better match human compositional patterns. We deploy the proposed framework as an Internet-based system that supports five distinct music styles and is publicly accessible online.